agent missed that the DPA's "Sub-processor" definition changed between the 2024 and 2026 versions

TL;DR

Diff the definitions section across DPA versions before extracting. Hash each defined term per version, re-extract every term whose hash changed, and pin the DPA version on every extraction. DPAs get versioned quietly and agents cache the first definition they saw, so the 2024 "Sub-processor" list keeps shipping after the 2026 version replaced it.

agent missed that the DPA's "Sub-processor" definition changed between the 2024 and 2026 versions

Steps

1. Index DPA versions explicitly (2024, 2026) instead of treating the DPA as one document. Expected: two versioned DPA records, each with its own definitions section

2. Hash each defined term's full text per version and diff. Expected: "Sub-processor" hash differs between 2024 and 2026; unchanged terms match

3. Re-extract only the drifted terms from the newer version. Expected: the sub-processor list is rebuilt from the 2026 definition

4. Show the before/after definition text for each drifted term in the report. Expected: the operator sees what the 2026 version changed

5. Pin the DPA version on every extraction that uses a DPA-defined term. Expected: citations read like '"Sub-processor" (DPA 2026)'

Use this when

• A DPA exists in multiple versions and defined terms may have changed.
• Sub-processor lists, data-retention terms, or transfer mechanisms look outdated.
• The agent indexed the DPA once and never re-checked it.
• You are auditing privacy-term extraction against the current DPA.

Not for this skill when

• There is only one DPA version in the bundle.
• The definition text is identical across versions; the drift is elsewhere.
• The sub-processor list lives outside the DPA (e.g. a public webpage); version the actual source instead.
• You need a legal assessment of whether the 2026 change is material.

Variant phrasings

DPA definition changed between versions, agent missed it

stale sub-processor list from old DPA version

"Sub-processor" definition drift 2024 to 2026

agent extracted DPA terms from superseded version

Why it happens

DPAs arrive as attachments and get parsed once at intake, usually the first version attached to the MSA. When the counterparty sends the 2026 DPA, it looks like a duplicate file, and dedup logic either drops it or files it without re-running definition extraction. The term cache is keyed by term name, not by (term, version), so the 2024 definition keeps winning every lookup.

Edge cases

• The 2026 DPA incorporates the 2024 definitions by reference except for listed changes. Parse the delta list; unchanged terms keep their 2024 text explicitly.
• The sub-processor list is maintained on a webpage the DPA links to. Version the fetched page with a fetch date; the DPA version alone is not enough.
• The bundle contains a draft 2026 DPA alongside the executed one. Only the executed version feeds extraction; mark the draft clearly.
• The definition changed but the operative list did not (or vice versa). Re-extract both; drift in one does not imply drift in the other.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstNI4cv86u6wK2O3ktbUoEw